December 2, 2024 2:30, ENR2 S215
When
Title: Multi-Fidelity Strategies for Reliability Assessment and Surrogate Construction
Abstract: Modern applications of uncertainty quantification and optimization methods are often
hampered by the large computational time of complex simulations (e.g., nonlinear
structural dynamics or computational fluid dynamics) and the potentially large
dimensionality of the problem. Multi-fidelity techniques, whereby the availability of a
hierarchy of models of various fidelity and cost are leveraged, are a popular way to
mitigate these limitations. This seminar will introduce two distinct active learning and
model management strategies for the calculation of probabilities of failure and the
construction of surrogates. The first approach is purely based on a classification
scheme and the computation of the probability of inconsistency between the class
predictions from the various models. This enables the formulation of an efficient
adaptive sampling strategy and model management. The second approach for the
construction of regression models is based on a hybrid combination of heteroscedastic
Gaussian processes and a Support Vector Machine classifier. The corresponding active
learning strategies and model management are both based on estimates of epistemic
and aleatoric uncertainty over the parameter space. The ability of both approaches to
reduce the number of high-fidelity calls is demonstrated on a series of problems of
varying dimensionality, including problems exhibiting discontinuities.
Relevant Papers: